A Beijing open-source model lands within 1 point of GPT-6 Astra on the AI tests that teach robots to act

A robot arm tries to slot a blue puzzle piece into a recess and fails. That single task, a one-centimetre insertion, is the field's standing bottleneck, including for the closed-source leader. This week a Beijing lab, DeepCybo, released PhysBrain 1.5, the first open-source model within a single point of GPT-6 Astra and Gemini 3.6 Flash on a 28-benchmark suite designed to teach machines to act on the real world. The 1-point gap is real. The open foundation is real. The precision gap that explains the puzzle-piece failure is also still real.

"Physical AI" is the term for models that have to do things in the world, not just describe it. A chatbot answers a question. A physical-AI model has to look at a camera feed, decide where a gripper should move, and predict what will happen next if it does. It is the difference between a model that can write a sentence about a kitchen and one that can clear a plate from a table. Researchers usually call this category embodied AI, and the benchmark suites that measure it draw on household robotics, drone navigation, and industrial manipulation. Most of the strongest models in this category are closed: their weights are private, their eval harness is private, and anyone who wants to build on them has to pay per call.

PhysBrain 1.5 is the new open entry, and its mechanism is the part that actually matters. DeepCybo trained it as a single unified vision-language model with a Qwen3-VL backbone and no task-specific heads. The model does everything through one autoregressive next-token objective, predicting the next token across language, structured spatial outputs, end-effector trajectories, and future world states. The same neural network that reads a sentence is also the one that proposes a gripper path and predicts what the scene will look like a second later. The lab published the technical report, both the 2-billion and 8-billion parameter weights, and a benchmark kit called PhysBrainEvalKit on the project page, with a public 8-billion-parameter demo on Hugging Face for anyone to test.

On the 28-benchmark suite, DeepCybo's release reports PhysBrain 1.5-8B at a 72.5 average, ranking first among open-source models. The model led on 14 of the 28 benchmarks and placed second on 10. The next open-source entries sit well behind: Hy-Embodied-VLM-1.0 at 66.0 and RynnBrain 1.1 at 63.1. DeepCybo's own 2-billion-parameter sibling scored 66.6, which the lab keeps out of the open-source ranking per its own rules even though the number tops every non-PhysBrain open entry on the leaderboard.

28-benchmark PhysBrainEvalKit average

All closed-model numbers on this suite come from DeepCybo's own runs, not from the closed-model vendors, so the 1-point gap is a one-party measurement until an independent re-run appears.

The same release names the capability buckets where the 8-billion-parameter model pulls ahead: visual-spatial perception, multi-view understanding, embodied cognition and planning, spatial affordance, and visual trajectory reasoning. Standout subscores include RoboSpatial-Home at 73.9, Part-Affordance at 84.0, and RoboRefit at 89.8. Those are the categories that drive household robotics and assembly, which is why the open weights land as a substrate event for downstream builders rather than a chatbot story.

Robocurve, an independent test on YAM arms under the Inspect Robots policy, put GPT-6 Astra at 19 of 20 on a place-the-block-in-the-bowl task versus Claude Fable 5.1's 8 of 20, at 2.5 minutes versus 6.8 minutes, and at roughly $0.94 versus $2.12 per run. On the same hardware, a millimetre-scale insertion task exposed the field-wide stall: GPT-6 Astra placed the part 2 of 20 times, the same as Fable 5.1, with both models stalling at the same final contact step. QbitAI's reporting frames the same result as a precision collapse from 95% to 10% on the hardest contact step. Either way, the closed-source leader is winning the macro task and still losing the contact-precision step.

That is the actual frontier the open weights have to attack. The puzzle-piece failure is not a benchmark quirk; it is the part of the loop where a model has to convert a learned spatial plan into a physical press, hold, and release, and the precision of that last step is what the next round of embodied research will measure. Google DeepMind released its own embodied-reasoning brain, Gemini Robotics 2, on July 30, and OpenAI has been positioning GPT-6 as a ceiling-raiser on reasoning and multimodal work. The 8-billion-parameter PhysBrain lands in that same week as a fully open entry, with weights, eval kit, and demo live.

The watch item is independent re-evaluation of the 28-benchmark suite. Until another lab re-runs PhysBrainEvalKit, the 1-point headline is a one-party number, and the only third-party robotic test in the public packet still shows the closed leader and the open challenger hitting the same precision wall. The bottleneck is moving, but it has not moved yet.